Using a Grammar Checker for Evaluation and Postprocessing of ...

Using a Grammar Checker for Evaluation and Postprocessing of Statistical
Machine Translation
Sara Stymne, Lars Ahrenberg
Department of Computer and Information Science
Linköping University, Sweden
{sarst,lah}@ida.liu.se
Abstract
One problem in statistical machine translation (SMT) is that the output often is ungrammatical. To address this issue, we have investigated
the use of a grammar checker for two purposes in connection with SMT: as an evaluation tool and as a postprocessing tool. As an
evaluation tool the grammar checker gives a complementary picture to standard metrics such as Bleu, which do not account for grammaticality.
We use the grammar checker as a postprocessing tool by applying the error correction suggestions it gives. There are only
small overall improvements of the postprocessing on automatic metrics, but the sentences that are affected by the changes are improved,
as shown both by automatic metrics and by a human error analysis. These results indicate that grammar checker techniques are a useful
complement to SMT.
1. Introduction
One problem with standard statistical machine translation
systems is that their output tends to be ungrammatical,
since there generally is no linguistic knowledge used in the
systems. We investigate how a grammar checker can be
used to address this issue. Grammar checkers are developed
to find errors in texts produced by humans, but in this
study we investigate if they can also be used to find errors
made by machines. We identify two novel usages of the
grammar checker for machine translation (MT): as an evaluation
tool and as a postprocessing tool.
We have performed experiments for English-Swedish
translation using a factored phrase-based statistical machine
translation (PBSMT) system based on Moses (Koehn
et al., 2007) and the mainly rule-based Swedish grammar
checker Granska (Domeij et al., 2000; Knutsson, 2001).
The combination of a grammar checker and a MT system
could be used for other architectures and language
pairs as well, however. We have performed experiments
on six translation systems that differ on two dimensions:
the amount of training data, and the amount of linguistic
knowledge used in the system.
To be able to use the grammar checker as an evaluation
tool, we performed an error analysis of the grammar
checker on SMT output. We then defined three crude
measures based on the error identification by the grammar
checker. All three measures are error rates based on the
grammar checker error categories. The difference between
them is that they use different subsets of the categories. All
three measures give a complementary picture to two standard
MT metrics, since they are better at accounting for the
fluency and grammaticality of the machine translation output.
We used the grammar checker as an automatic postprocessing
tool on the SMT output, by using the correction
suggestions given for many errors. We applied the suggestions
only for categories that had a high precision on
the error analysis on SMT output. A human error analysis
showed that the corrections were successful in most cases.
There were only small improvements on automatic metrics
on the full test sets, however, but this was mainly due to the
fact that the postprocessing only affected a small proportion
of the sentences.
The remainder of the paper is organized as follows: Section
2 describes the translation systems used in the study. In
Section 3 the grammar checker Granska is introduced, and
an error analysis of Granska on SMT output is described.
The experiments and results are presented in Section 4 for
evaluation and in Section 5 for postprocessing. Section 6
contains a discussion of related work, and Section 7 contains
the conclusion and suggestions for future work.
2. SMT System
The translation system used is a standard PBSMT setup using
the Moses decoder (Koehn et al., 2007) and the SRILM
toolkit for sequence models (Stolcke, 2002). We take advantage
of the factored translation framework in Moses
(Koehn and Hoang, 2007), where factors other than surface
form can be used to represent words, which allows the
inclusion of linguistic knowledge such as lemmas and partof-speech
tags. To tune feature weights minimum error rate
training is used (Och, 2003).
The system is trained and tested on the Europarl corpus
(Koehn, 2005). The Swedish target side of the training
corpus is part-of-speech tagged using the Granska tagger
(Carlberger and Kann, 1999). The training corpus is filtered
to remove sentences longer than 40 words and with a
length ratio of more than 1 to 7. 500 sentences are used for
tuning and 2000 sentences for testing.
In order to evaluate the use of the grammar checker we
trained six different systems that are varied on two dimensions:
the amount of training data and the amount of linguistic
knowledge in the systems, in the form of extra factors
on the output side. For the corpus size we have used
two sizes of the training corpus for the translation model:
a large training corpus with 701,157 sentences, and a small
training corpus with 100,000 sentences. In both cases the
sequence models were trained on the large corpus.

Text: Averaging vore med tre timmar per dag , det är den mest omfattande mänskliga
aktivitet efter sover och - för vuxna - arbete .
Rule: stav1@stavning Span: 1-1 Words: Averaging
Rule: kong10E@kong Span: 14-15 Words: mänskliga aktivitet
mänskliga aktiviteten
mänsklig aktivitet
Figure 1: Example of filtered Granska output for a sentence with two errors
The linguistic knowledge used in the system are varied
by the use of different factors on the output side. There are
three system setups for each corpus size: the none system
with only a standard language model on surface form, and
two systems with an additional sequence model. The POS
system use standard part-of-speech tags and the morph system
use morphologically enriched part-of-speech tags. An
example of the annotation is shown in (1).
(1) EN: my question is important .
none: min fråga är viktig .
POS: min|ps fråga|nn är|vb viktig|jj .|mad
morph: min|ps.utr.sin.def fråga|nn.utr.sin.ind.nom
är|vb.prs.akt.kop viktig|jj.pos.utr.sin.ind.nom .|mad
Using part-of-speech tags can be expected to improve word
order. The use of morphological tags can also improve
agreement (Stymne et al., 2008).
3. Grammar Checker
A grammar checker is a tool that can identify grammar errors
in texts. Often they also include other errors such as
spelling errors and stylistic errors. Grammar checkers tend
to be authoring tools or writing aids, that is, they are designed
to be used by a human who assess the alarms and
suggestions given by the tools, rather than as software that
can be applied automatically.
For MT output, both types of grammar checkers could be
considered useful. A grammar checker system that is an authoring
tool could be used to highlight suspicious problems
and to suggest changes in translations that are sent to a human
posteditor. An automatic grammar checking system,
on the other hand, could be used to automatically improve
MT output, regardless of whether the translations are being
used directly, e.g. for gisting, or if they are being sent for
further postediting by humans. If the purpose of using the
grammar checker is evaluation it would clearly be preferable
with an automatic grammar checker.
3.1. Granska
We use the Swedish grammar checker Granska (Domeij et
al., 2000; Knutsson, 2001), which is a hybrid, mainly rulebased
grammar checker. The main modules in Granska are:
a tokenizer; a probabilistic Hidden Markov model-based
tagger, which tags texts both with part-of-speech and morphology
(Carlberger and Kann, 1999); the spell checker
Stava (Kann et al., 2001); and a rule matcher, which identifies
errors and generates error descriptions and correction
suggestions. The rule matcher contains hand-written rules
in an object-oriented rule language developed for Granska.
Granska finds both grammar and spelling errors, and in
many cases it also gives correction suggestions. The grammar
errors are divided into thirteen main categories, of
which many are in turn further divided into a number of
subcategories.
Granska can output XML, from which we extracted the
necessary information for our purposes, exemplified in Figure
1. In the output we see the whole sentence (Text:), and
a list of errors that were found in each sentence. For each
error we know the main rule type and rule subcategory that
applied (Rule:), the position in the sentence where it occurs
(Span:), the words that are wrong (Words:), and possibly
one or several correction suggestions. For the sentence in
Figure 1, Granska detected a spelling error, the unknown
English word Averaging, for which it has no correction suggestions,
and a noun phrase agreement error, for which it
has two suggestions, of which the first one is correct in this
context. There are also other errors in the sentence which
Granska does not find, mainly because it is not designed for
this type of malformed output.
3.2. Error Analysis of Granska on SMT Output
In order to use the grammar checker for SMT it is useful to
know on which categories the grammar checker produces
good results and good correction suggestions for SMT output.
Granska has been evaluated on human output in previous
studies, but with different results on different text types
(Knutsson, 2001). Applying Granska on a different type of
text than it was designed for can also affect its performance;
for instance, its precision for a subset of its error categories
degrades from 92% on texts written by adults to 35% on
texts written by primary school children (Sofkova Hashemi,
2007). Machine translation output is another very different
text type, on which we cannot expect the behavior of
Granska to be the same as for human texts. Thus we performed
an error analysis of the performance of the grammar
checker on the translation output from the tuning process,
a total of 11,574 words, from Europarl. The evaluation was
performed by one native Swedish speaker.
The grammar errors found in the sample only belonged
to five of the thirteen error categories in Granska:
• Agreement NP – Errors in noun phrase agreement of
determiners, nouns and adjectives for number, gender
and definiteness
• Agreement predicatives – Errors in agreement of
predicatives with the subject or object on number, gender
and definiteness

Type Error identification Correction suggestions
Correct False Correct1 Correct2+ Wrong None
Agreement NP 64 10 48 10 4+10 2+0
Agreement Predicatives 21 1 20 – 1+1 –
Split compounds 12 14 8 – 3+13 1+1
Verb 31 18 11 2 – 18+18
Word order 9 0 8 – 1+0 –
Table 1: Analysis of grammar errors that are identified by Granska on tuning data. Correct1 means that the top suggestion
is correct, and Correct2+ that some other suggestion is correct. For cases where suggestions are wrong or not given, the
numbers are divided between errors that are correctly identified and errors that are erroneously identified.
Type Granska evaluation on human text Granska evaluation on SMT output
Recall Precision Precision range Precision
Agreement NP 0.83 0.44 0.11–0.72 0.86
Agreement Predicative 0.69 0.32 0.00–0.44 0.95
Split compounds 0.46 0.39 0.00–0.67 0.46
Verb 0.97 0.83 0.71–0.91 0.63
Word order 0.75 0.38 0.00–1.00 1.00
Table 2: Comparison of an evaluation of Granska on human text (Knutsson, 2001) with the error analysis on SMT output.
Precision range is the extreme precision values on the five different text types
• Split compounds – Compounds that are written as
two separate words, instead of as one word
• Verb – Errors that are related to verbs, such as missing
verbs, wrong form of verbs, or two finite verbs
• Word order – Wrong word order
Table 1 summarizes the results for these five categories.
The performance varies a lot between the error categories,
with good performance of error identification on agreement
and word order, but worse on split compounds and
verbs. Looking into the different subcategories shows that
the false alarms for verbs mostly belong to three categories:
missing verb, missing finite verb, and infinitive marker
without a verb. When we excluded these categories there
are 17 verb errors left, of which only 1 is wrong.
The quality of the correction suggestions also varies between
the categories. In the verb category, suggestions are
never wrong, but they are given in a minority of the cases
where they are correctly identified, and never for the false
alarms. For split compounds, on the other hand, the majority
of the correction suggestions are incorrect, mainly
since suggestions are given to nearly all false alarms. For
NP agreement, all false alarms have correction suggestions,
but the majority of the correction suggestions are still correct.
Predicative agreement and verb errors have correct
suggestions for nearly all identified errors.
There are very few pure spelling errors in the translation
output, since the words all come from the corpus the
SMT system is trained on. The grammar checker still identifies
161 spelling errors, of which the majority are untranslated
foreign words (49.0%) and proper names (32.9%).
The only correction suggestions that are useful for spelling
errors is the capitalization of lower-cased proper names,
which occur in 9 cases.
The error analysis on SMT output can be contrasted to
an earlier evaluation of Granska on human texts performed
by Knutsson (2001). That evaluation was performed on
201,019 words from five different text types: sport news,
foreign news, government texts, popular science and student
essays. The results from that evaluation on the error
categories found in the SMT sample are contrasted with the
SMT error analysis in Table 2. The performance of Granska
varies a lot between the five text types, as shown by the precision
range. The precision on SMT output is better than
the average precision on human texts on all error categories
except verb errors. This is promising for the use of Granska
on SMT output. The human texts were annotated with all
present errors, which meant that recall could be calculated.
The recall is rather high on all categories except split compounds.
No annotation of all errors were done in our SMT
error analysis, and thus we cannot calculate recall. It could
be expected to be a lot lower than on human texts, however,
since Granska was not developed with SMT errors in mind.
4. Grammar Checker as an Evaluation Tool
A disadvantage of most current evaluation metrics for MT,
is that they do not take grammar into account. Thus, the
grammar checker could complement existing metrics by indicating
the grammaticality of a text, simply by counting
the number of errors. This only accounts for the fluency of
the output, however, so it needs to be used in addition to
other metrics that can account for adequacy. In this study
we compare the translation results on three crude measures
based on the grammar checker with the results on two standard
metrics: Bleu (Papineni et al., 2002), which mainly
measures n-gram precision, and TER (Snover et al., 2006),
which is an error rate based on a modified Levenshtein distance.
1 The scores for all metrics are calculated based on a
single reference translation.
Based on the error analysis in Section 3.2. we define
three grammar checker based metrics. Grammar error ratio
1 Percent notation is used for all Bleu and TER scores.

Size Factors Bleu TER GER 1 GER 2 SGER
none 22.18 66.42 0.196 0.293 0.496
Large POS 21.63 66.88 0.228 0.304 0.559
morph 22.04 66.63 0.125 0.195 0.446
none 21.16 67.49 0.244 0.359 0.664
Small POS 20.79 67.79 0.282 0.375 0.718
morph 19.45 69.52 0.121 0.245 0.600
Table 3: Results for the six basic systems. Bleu have higher values for better systems, whereas TER and the grammar
checker metrics are error rates, and have lower values for better systems.
1, GER 1 , is the average number of the grammar errors that
had a high precision on error categorization on the development
set (all errors except split compounds and three verb
categories), and grammar error ratio 2, GER 2 , is the average
number of all grammar errors per sentence. Spelling
and grammar error ratio, SGER, is the average total number
of identified errors per sentence. All three metrics are error
rates, with the value of 0 for systems with no identified
errors, and a higher value when more errors are identified.
There is no upper bound on the metrics.
Table 3 shows the results of the evaluation. As expected
the systems trained on the large corpus are all better than
the small systems on the standard metrics. On the new
grammar checker metrics the large systems are all better
on SGER but not always on the GER metrics. The ranking
between the six systems are the same on the two standard
metrics, but differ for the Granska metrics. The morph
system is best for the small corpus on the Granska metrics,
but markedly worse on the standard metrics. We believe
that this is because the small corpus is too small for
the morphology to be useful, due to sparsity issues, so that
the grammar is improved, but the word selection is worse.
Also on the large corpus, the morph sequence model does
improve the grammaticality of the output, but here it performs
nearly equal to the none system, and better than the
POS system on the standard metrics. Using a POS sequence
model gives the worst performance on all metrics for both
corpus sizes.
There are some interesting differences between the three
Granska metrics. SGER gives markedly worse performance
on the small corpus, which is not the case for the
two GER metrics. This is mainly due to the fact that SGER
is the only metric that includes spelling errors, and many
of the spelling errors are English out-of-vocabulary words,
which are just passed through the translation system. With
a smaller training corpus, the number of out-of-vocabulary
words increases. If we want the metric to give a picture
of the coverage of the system, this information is clearly
useful. GER 1 gives the best value for the morph system
with the small corpus. This could be an indication that using
only these few error categories makes this metric less
robust to bad translations, since the grammar checker performance
most likely degrades when the SMT output is too
bad for it to analyze in a meaningful way. On the other hand
GER 1 gives a larger difference between the none and POS
systems, than GER 2 , which also uses error categories with
a bad performance on SMT output.
Size Factors Bleu TER Changes
none 22.34 66.37 382
Large POS 21.81 66.84 429
morph 22.17 66.54 259
none 21.30 67.47 456
Small POS 20.95 67.75 514
morph 19.52 69.48 249
Table 4: Results and number of changes for the six systems
when Granska is used for postprocesing
5. Grammar Checker for Postprocessing
We wanted to see if the suggestions of the grammar checker
could be used to improve MT by automatic postprocessing,
where we apply the suggestions from the grammar checker.
We have chosen to accept the correction suggestions for the
categories where a large majority of the suggestion were
correct in the error analysis in Section 3.2. In the test systems
there were between 8 and 16 errors in categories that
did not appear in the error analysis. These errors were ignored.
When there were several correction suggestions for
an error, we always chose the first suggestion. For most of
the errors, such as agreement errors, the corrections were
performed by changing one or several word forms. For
other categories, the word order was changed, or words
were deleted or inserted.
Table 4 shows the results for the systems with postprocessing.
There are consistent but very small improvements
for all systems on both metrics, compared to the unprocessed
scores in Table 3. The Bleu scores had an absolute
improvement by at most 0.18, and the TER scores by just
0.09 at most. As could be expected, the systems with bad
scores on the Granska metrics have more corrections. More
interestingly, the improvements as measured by the metrics
were not generally larger on the none and POS systems than
on the morph systems with much fewer corrections.
One of the likely reasons for the small improvements on
the two metrics was that only a relatively small proportion
of the sentences were affected by the postprocessing. To
investigate this, we calculated scores on only the subset of
sentences that were affected by the changes, shown in Table
5. The subsets of sentences are all different, so the scores
can not be directly compared between the systems, or to the
scores on the full test sets. Not surprisingly, the improvements
are much larger on the subsets than on the full test set
for all systems. The difference in the change is much larger
on Bleu, with an absolute improvement of around 0.70 for
the large systems and of around 0.50 for the small systems.

Size Factors Bleu TER No. sentences
Basic Postproc. Basic Postproc.
none 19.44 20.12 68.99 68.77 335
Large POS 18.87 19.61 69.42 69.26 373
morph 18.47 19.29 70.28 69.69 238
none 18.72 19.26 69.96 69.88 395
Small POS 17.74 18.27 71.00 70.88 452
morph 16.79 17.24 72.01 71.77 241
Table 5: Results on the subsets of sentences where Granska postprocessing led to a change. Note that these subsets are
different for the six systems, so these scores are not comparable between the systems.
Size Factors Good Neutral Bad
none 73 19 8
Large POS 77 17 6
morph 68 19 13
none 74 19 7
Small POS 73 17 10
morph 68 20 12
Table 6: Error analysis of the 100 first Granska-based
changes for each system
The difference on TER is still relatively small, except for
the morph systems, especially for the large system, which
has a TER improvement of 0.59.
To further investigate the quality of the error corrections,
an error analysis was performed by one native Swedish
speaker on the first 100 error corrections for each system.
The corrections were classified into three categories: good,
bad, and neutral. The good changes were improvements
compared to not changing anything, and the bad changes resulted
in a worse translation than before the change. For the
neutral category the translations before and after the change
were of equal quality.
The results of the analysis are summarized in Table 6. A
large majority of the changes were improvements, which
indicates that the two standard metrics used are not that
good in capturing this type of improvement. The corrections
for the morph systems, are slightly worse than for the
other systems. This is to a large degree due to the fact that
these systems have fewer agreement errors than the other
systems, and agreement errors are generally the easiest errors
to correct.
Table 7 shows some examples of the different types of
errors. In the good example a NP agreement error is fixed
by switching the indefinite article so it gets the correct gender,
and a verb with the wrong tense, present, is changed
to perfect tense. In the first neutral example an NP agreement
error which is mixed between definite and indefinite,
is changed and becomes syntactically correct, but indefinite
instead of the preferred definite. The second neutral
example concerns agreement of an adjectival predicative
with a collective noun, where both the original plural adjective,
and the changed singular adjective are acceptable.
In the first bad example an attributive adjective has been
mistaken for a head noun, resulting in a change from a correct
NP to an incorrect NP with two noun forms. The second
bad example contains an untranslated English plural
genitive noun, which are given Swedish plural inflection by
Granska.
Even though the performed corrections are generally
good, the number of errors with useful suggestions is low,
which makes the overall effect of the corrections small.
There are many more actual errors in the output, which are
not found by Granska. For the postprocessing technique to
be even more useful we need to be able to identify and correct
more of the errors, either by modifying the grammar
checker or by developing a custom SMT output checker.
6. Related Work
Automatic metrics are usually based on the matching of
words in the translation hypothesis to words in one or several
human reference translations in some way, as is done
by Bleu (Papineni et al., 2002) and TER (Snover et al.,
2006). These types of metrics do not generalize any linguistic
knowledge; they only rely on the matching of surface
strings. There are metrics with extended matching,
e.g. Meteor (Lavie and Agarwal, 2007), which uses stemming
and synonyms from WordNet, and TERp (Snover et
al., 2009), which uses paraphrases. There are also some
metrics that incorporate other linguistic levels, such as partof-speech
(Popović and Ney, 2009), dependency structures
(Owczarzak et al., 2007), or deeper linguistic knowledge
such as semantic roles and discourse representation structure
(Giménez and Márquez, 2008).
Controlled language checkers, which can be viewed as
a type of grammar checker, have been suggested in connection
to MT, but for preprocessing of the source language,
see Nyberg et al. (2003) for an overview. The
controlled language checkers tend to be authoring tools,
as in Mitamura (1999) and de Koning (1996) for English
and Sågvall Hein (1997) for Swedish, which are used by
humans before feeding a text to a usually rule-based MT
system.
Automatic postprocessing has been suggested before for
MT, but not by using a grammar checker. Often postprocessing
has targeted specific phenomena, such as correcting
English determiners (Knight and Chander, 1994), merging
German compounds (Stymne, 2009), or applying word
substitution (Elming, 2006). The combination of a statistical
MT system and the mainly rule-based grammar checker
can also be viewed as a hybrid MT system, on which there
has been much research, see e.g., Thurmair (2009) for an
overview. Carter and Monz (2009) discuss the issue of applying
tools that are developed for human texts, in their case

Good
Neutral
Bad
Original
Changed
Original
Changed
Correct
Att ge nya befogenheter till en kommitté av ministrar främjas genom en oansvarigt sekretariat
skulle inte utgör någon typ av framsteg . . .
Att ge nya befogenheter till en kommitté av ministrar främjas genom ett oansvarigt sekretariat
skulle inte ha utgjort någon typ av framsteg . . .
Det är viktigt att fylla den kulturella vakuum mellan våra två regioner
Det är viktigt att fylla ett kulturellt vakuum mellan våra två regioner
Det är viktigt att fylla det kulturella vakuumet mellan våra två regioner
Original Jag hör ibland sägas att rådet är så engagerade i Berlin . . .
Changed Jag hör ibland sägas att rådet är så engagerat i Berlin . . .
Original Skulle det inte vara värt att ansvar på alla nivåer i den beslutsfattande processen tydligare,
snarare än att försöka gå framåt . . .
Changed Skulle det inte vara värt att ansvar på alla nivåer i det beslutsfattandet processen tydligare,
snarare än att försöka gå framåt . . .
Original Dokumentet kommer att överlämnas till europeiska rådet i Biarritz i några days’ tid .
Changed Dokumentet kommer att överlämnas till europeiska rådet i Biarritz i några daysar tid .
Correct Dokumentet kommer att överlämnas till europeiska rådet i Biarritz i några dagars tid .
Table 7: Some examples of error corrections from the different categories
statistical parsers, on SMT output.
7. Conclusion and Future Work
We have explored the use of a grammar checker for MT, and
shown that it can be useful both for evaluation and for postprocessing.
A more large scale investigation with a more
thorough error analysis would be useful, both for Swedish
with the Granska grammar checker and for other languages
and tools. In particular we want to see if the results are
as useful for other architectures of the MT system and of
the grammar checker, as for the combination of a statistical
MT system and a mainly rule-based grammar checker. We
also plan to apply grammar checkers for other languages
on standard datasets such as that of the WMT shared task 2
(Callison-Burch et al., 2009), where we could correlate the
grammar checker performance with human judgments and
several automatic metrics for many different MT systems.
Using the grammar checker for evaluation gives a complimentary
picture to the Bleu and TER metrics, since
Granska accounts for fluency to a higher extent. Granska
needs to be combined with some other metric to account for
adequacy, however. A possibility for future research would
be to combine a grammar checker and some measure of adequacy
into one metric.
In the postprocessing scenario we used the grammar
checker as a black box with good results. The grammar
checker was, however, only able to find a small proportion
of all errors present in the SMT output, since it was not developed
with SMT in mind. One possibility to find more
errors is to extend the rules in the grammar checker. Another
possibility would be a tighter integration between the
SMT system and the grammar checker. As an example, the
grammar checker tags the translation output, which is errorprone.
Instead part-of-speech tags from factored translation
systems could be used directly by a grammar checker, without
re-tagging. A third option would be to develop a new
2 The shared translation task at the Workshop of Statistical
Machine Translation, http://www.statmt.org/wmt09/
translation-task.html
grammar checker targeted at MT errors rather than human
errors, which could be either a stand-alone tool or a module
in a MT system.
8. References
Chris Callison-Burch, Philipp Koehn, Christof Monz, and
Josh Schroeder. 2009. Findings of the 2009 Workshop
on Statistical Machine Translation. In Proceedings of
the Fourth Workshop on Statistical Machine Translation,
pages 1–28, Athens, Greece.
Johan Carlberger and Viggo Kann. 1999. Implementing
an efficient part-of-speech tagger. Software Practice and
Experience, 29:815–832.
Simon Carter and Christof Monz. 2009. Parsing statistical
machine translation output. In Proceedings of the 4th
Language & Technology Conference, Poznań , Poland.
Michiel de Koning. 1996. Bringing controlled language
support to the desktop. In Proceedings of the EAMT Machine
Translation Workshop, pages 11–20, Vienna, Austria.
Rickard Domeij, Ola Knutsson, Johan Carlberger, and
Viggo Kann. 2000. Granska – an efficient hybrid system
for Swedish grammar checking. In Proceedings of
the 12th Nordic Conference on Computational Linguistics
(Nodalida’99), pages 49–56, Trondheim, Norway.
Jakob Elming. 2006. Transformation-based correction of
rule-based MT. In Proceedings of the 11th Annual Conference
of the European Association for Machine Translation,
pages 219–226, Oslo, Norway.
Jesús Giménez and Lluìs Márquez. 2008. A smorgasbord
of features for automatic MT evaluation. In Proceedings
of the Third Workshop on Statistical Machine Translation,
pages 195–198, Columbus, Ohio, USA.
Viggo Kann, Rickard Domeij, Joachim Hollman, and
Mikael Tillenius. 2001. Implementation aspects and applications
of a spelling correction algorithm. Quantitative
Linguistics, 60:108–123.
Kevin Knight and Ishwar Chander. 1994. Automated
postediting of documents. In 12th National conference

of the American Association for Artificial Intelligence,
Seattle, Washington, USA.
Ola Knutsson. 2001. Automatisk språkgranskning av
svensk text. Licentiate thesis, Royal Institute of Technology
(KTH), Stockholm, Sweden.
Philipp Koehn and Hieu Hoang. 2007. Factored translation
models. In Proceedings of the 2007 Joint Conference on
Empirical Methods in Natural language Processing and
Computational Natural Language Learning, pages 868–
876, Prague, Czech Republic.
Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris
Callison-Burch, Marcello Federico, Nicola Bertoldi,
Brooke Cowan, Wade Shen, Christine Moran, Richard
Zens, Chris Dyer, Ondrej Bojar, Alexandra Constantin,
and Evan Herbst. 2007. Moses: Open source toolkit
for statistical machine translation. In Proceedings of the
45th Annual Meeting of the ACL, demonstration session,
pages 177–180, Prague, Czech Republic.
Philipp Koehn. 2005. Europarl: A parallel corpus for statistical
machine translation. In Proceedings of MT Summit
X, pages 79–86, Phuket, Thailand.
Alon Lavie and Abhaya Agarwal. 2007. METEOR: An
automatic metric for MT evaluation with high levels of
correlation with human judgments. In Proceedings of
the Second Workshop on Statistical Machine Translation,
pages 228–231, Prague, Czech Republic.
Teruko Mitamura. 1999. Controlled language for multilingual
machine translation. In Proceedings of MT Summit
VII, pages 46–52, Singapore, Singapore.
Eric Nyberg, Teruko Mitamura, and Willem-Olaf Huijsen.
2003. Controlled language for authoring and translation.
In Harold Somers, editor, Computers and Translation.
A translator’s guide, pages 245–282. John Benjamins,
Amsterdam.
Franz Josef Och. 2003. Minimum error rate training in statistical
machine translation. In Proceedings of the 42nd
Annual Meeting of the ACL, pages 160–167, Sapporo,
Japan.
Karolina Owczarzak, Josef van Genabith, and Andy Way.
2007. Dependency-based automatic evaluation for machine
translation. In Proceedings of the Workshop on
Syntax and Structure in Statistical Translation, pages
80–87, Rochester, New York, USA.
Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing
Zhu. 2002. BLEU: A method for automatic evaluation
of machine translation. In Proceedings of the 40th Annual
Meeting of the ACL, pages 311–318, Philadelphia,
Pennsylvania, USA.
Maja Popović and Hermann Ney. 2009. Syntax-oriented
evaluation measures for machine translation output. In
Proceedings of the Fourth Workshop on Statistical Machine
Translation, pages 29–32, Athens, Greece.
Anna Sågvall Hein. 1997. Language control and machine
translation. In Proceedings of the 7th International Conference
of Theoretical and Methodological Issues in Machine
Translation, pages 103–110, Santa Fe, New Mexico,
USA.
Matthew Snover, Bonnie Dorr, Richard Schwartz, Linnea
Micciulla, and John Makhoul. 2006. A study of translation
edit rate with targeted human notation. In Proceedings
of the 7th Conference of the Association for Machine
Translation in the Americas, pages 223–231, Cambridge,
Massachusetts, USA.
Matthew Snover, Nitin Madnani, Bonnie Dorr, and Richard
Schwartz. 2009. Fluency, adequacy, or HTER? Exploring
different human judgments with a tunable MT metric.
In Proceedings of the Fourth Workshop on Statistical
Machine Translation, pages 259–268, Athens, Greece.
Sylvana Sofkova Hashemi. 2007. Ambiguity resolution by
reordering rules in text containing errors. In Proceedings
of 10th International Conference on Parsing Technologies,
pages 69–79, Prague, Czech Republic.
Andreas Stolcke. 2002. SRILM – an extensible language
modeling toolkit. In Proceedings of the Seventh International
Conference on Spoken Language Processing,
pages 901–904, Denver, Colorado, USA.
Sara Stymne, Maria Holmqvist, and Lars Ahrenberg. 2008.
Effects of morphological analysis in translation between
German and English. In Proceedings of the Third Workshop
on Statistical Machine Translation, pages 135–138,
Columbus, Ohio, USA.
Sara Stymne. 2009. A comparison of merging strategies
for translation of German compounds. In Proceedings of
the EACL 2009 Student Research Workshop, pages 61–
69, Athens, Greece.
Gregor Thurmair. 2009. Comparing different architectures
of hybrid machine translation systems. In Proceedings
of MT Summit XII, pages 340–347, Ottawa, Ontario,
Canada.